Hey friend!

Welcome to the new Unpacking Meaning look. Let me know what you think, but I wanted to create a cohesive style with me new website I just launched (this email design was made by my AI agent Travis too btw). Anyay, on to today's topic...

A synthetic customer will never (for now) cancel your interview, misunderstand your questions, or tell you your new positioning makes no sense.

That is exactly the problem.

Over the past year, I’ve watched synthetic research move from an interesting experiment to something a marketing team can run regularly. You can create a panel, show it twenty headlines, ask follow-up questions, and get a tidy report in an afternoon. That speed helps, as long as you don’t mistake the report for actual evidence.

When I published The state of synthetic research in 2025, my conclusion was that the best approach was hybrid: use synthetic methods for early, directional exploration, then use real people and real behavior for validation. Or even better, use a hybrid of synthetic data rotted in human data.

Two recent sources point the same way. A May 2026 psychology preprint argued that large language models remain limited substitutes for human participants. In April, Greenbook’s industry guidance recommended synthetic research for pre-testing and directional exploration, with human signal kept in the loop for the decisions that matter.

That is also how I’m approaching an internal system I’m building to pressure-test messaging before it goes in front of customers. It creates simulated buyer roles, gives each one the same page and context, then checks which parts of the intended message survive a quick scan (btw, thanks to Expected Parrot for this). The output is a list of possible comprehension problems for a human to verify, not invented buyer quotes and not a replacement for real research.

Working on that system has clarified the practical boundary for me.

3 things synthetic research is good for

1. Generating hypotheses. Use it to surface questions, objections, or possible angles you may have missed, then investigate them properly.
2. Exploring bounded scenarios. Give the system a defined audience, context, and decision so you can examine possibilities before spending real budget.
3. Stress-testing and shortlisting. Run twenty messages, concepts, or research questions through it to find the five worth putting in front of people.

3 things it is not good for

1. Claiming customer truth. A model’s pattern-match is not observed evidence of what customers actually thought, said, or did.
2. Replacing discovery. Live conversations surface contradiction, emotion, culture, lived context, and answers you did not know to ask for.
3. Predicting high-stakes outcomes with precision. A score becomes decision-grade only after it has been calibrated against real customer evidence and behavior.

Going back to my touring days, this is exactly what we used our rehearsal room for with the band: to try out our live show, setlist and sound before performing in public.

Knowing this boundary should also change how you write up synthetic findings.

Instead of reporting “This audience prefers option A,” think: “Option A passed our synthetic stress test. Here is what we need to validate with customers next.”

Your result should open the right next question.

Speaking of rehearsals… what if your message has to persuade an agent?

The same rehearsal-versus-reality boundary applies to a newer messaging problem: what happens when an AI agent encounters your message before the customer does?

As people delegate more research and buying decisions to agents, those agents may become readers, critics, and gatekeepers between your product and its audience. Marketing teams will need to know whether an agent understands what the product is, who it is for, what evidence supports it, and why it is worth recommending to a human.

Synthetic research can help us rehearse that message. Ethan Mollick recently built one page for human readers and another specifically for AI agents to promote his new book. The agent-facing version makes a transparent case for why the book could help the human, supplies the facts an agent may need, offers a suggested message to relay, and asks the agent to get permission before starting a purchase.

In his account of the experiment, Mollick says he showed the pitch to multiple AI models, tested it repeatedly for different potential users, compared versions and file formats, and changed language the models interpreted as a suspicious instruction. In other words, he used AI to test messaging aimed at AI.

That suggests a useful new job for synthetic research: test whether agents can understand, evaluate, and accurately relay your message before you publish it.

The test should ask:

1. Can the agent explain what this is and who it’s for?
2. Can it distinguish the evidence from the marketing claim?
3. Does it know when the product is relevant and when it’s not?
4. Can it suggest a sensible next step without overstepping the customer’s permission?

This still does not tell us whether customers will accept the recommendation or buy. It tells us whether the message survives contact with the machines increasingly helping them decide.

I’m working on a 2026 refresh of the guide now. Whether the audience is human or agent, the discipline is the same:

Use simulation to find what deserves a real test—not to declare that the test has already passed.

Discovery

These two pieces helped me think about where agents belong, where taste belongs, and where a human gate still matters.

The self-driving company—and the marketing version we’re building toward

Amjad Masad’s The Self-Driving Company describes Replit giving employees a manager agent that coordinates specialists, checks results, and escalates when human judgment is needed.

That is close to an operating model I’m thinking about and building for developing for marketing teams: each lead works through an orchestrator that coordinates specialist agents for research, strategy, copy, and quality checks. The agents do more of the execution, while people provide the source material, approve the decisions, and own the consequences.

Normal is forgotten. Only weird survives.

George Mack’s essay suggests that when you are unsure what to do, write down five things most people would do in the same situation, cross them out, and try again.

Applied to messaging, the question might be: “How would every other company in our category explain this?” Write down the five predictable answers, then look for one that could only come from your experience and point of view. A simulation will happily return the category average. The taste and odd detail that change the message still have to come from people.

Resonance